{
  "id": 2183960,
  "title": "Up to 3.2x Faster Inference with LFM2.5-DSpark",
  "url": "https://urgent.news/2026/08/20/up-to-3-2x-faster-inference-with-lfm2-5-dspark",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-20T16:52:57.000Z",
  "source": {
    "name": "Hugging Face",
    "slug": "hugging-face",
    "url": "https://huggingface.co/blog/LiquidAI/lfm25-dspark"
  },
  "original_language": "en",
  "account": "Liquid AI has released draft model checkpoints for three models from their LFM2.5 family: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. These checkpoints introduce a speculative decoding path that enhances decoding speed without compromising output quality. The decoding phase in large language model (LLM) inference is traditionally memory-bound, with most latency coming from streaming weights from DRAM into SRAM. Speculative decoding addresses this issue by using a lightweight draft model to generate candidate tokens, which are then verified by the target model in a single forward pass. This approach shares the cost of loading weights across all verified tokens.\n\nThe DSpark model incorporates three components: a draft model, a lightweight attention-only version, and a verification stage. The draft models are based on simplified attention-only architectures with 5 layers and a block of 9. Each draft model undergoes 15 epochs on the entire dataset, with the epoch showing the highest acceptance rate selected. The resulting draft models are relatively small, with each containing approximately 300 million parameters.\n\nThe DSpark draft models for LFM2.5 are compatible with llama.cpp and SGLang, two open-source implementations. Benchmarks were conducted on-device using llama.cpp and Metal on an M4 Max MacBook Pro with FP16 GGUF weights and up to 256 output tokens, as well as on a single H100 80 GB GPU in BF16. Both configurations used a DSpark block size of 9, a batch size of 1, and a temperature of 0. The models were evaluated on five benchmark datasets, with all three draft models demonstrating significant throughput improvements on both large-scale and edge deployment platforms. For LFM2.5-2.6B, the MacBook configuration achieved an especially notable speedup, pushing the interactivity level far beyond what most proprietary cloud models offer. Across various multi-tool scenarios, DSpark reduced the latency by an average of 57% for LFM2.5-2.6B. The speedup for LFM2.5-1.2B-Instruct varied depending on the underlying text distribution, ranging from 0% to 52%. For LFM2.5-8B-A1B, the acceptance rate increased compared to two dense models, but on-device performance saw only an 18% improvement on average. This difference can be attributed to the current MoE implementation in llama.cpp's Metal backend and the increased weight traffic resulting from verifying multiple tokens simultaneously.",
  "summary": null,
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}